Predicting Online Peer-to-Peer(P2P) Lending Default using Data Mining Techniques

Transcription

1 Predicting Online Peer-to-Peer(P2P) Lending Default using Data Mining Techniques Jae Kwon Bae, Dept. of Management Information Systems, Keimyung University, Republic of Korea. Seung Yeon Lee and Hee Jin Seo, Dept. of Statistics, School of Business Keimyung University, Republic of Korea. Abstract In this study, we developed online Peer-to-Peer(P2P) lending default prediction models using logistic regression, decision trees (i.e., CART and C5.0), and multilayer perceptron (i.e., MLP) in order to predict P2P loan default. To verify the feasibility and effectiveness of P2P lending default prediction models, borrower loan data and credit data provided by Lending Club, the biggest United States P2P lending company used in this study. Empirical results indicated that MLP outperforms other classifiers such as logistic regression, CART, and C5.0. MLP always outperforms other classifiers in P2P loan default prediction. Key Words: Online Peer-to-Peer(P2P) lending, P2P loan default prediction, decision trees, multilayer perceptron, Lending Club JEL Classification: C 19, G13, G 14 1

2 1. Introduction Proceedings of the 20th Asia-Pacific Conference on Global Business, Economics, In recent years, online Peer-to-Peer (P2P) lending has developed rapidly in the world. P2P lending is to borrow and lend on the internet, and borrowers and lenders can use the internet platform without the intermediation of a financial institution. Since P2P lending companies (platform) offering these services generally operate online, they can run with lower overhead and provide the service more cheaply than traditional financial institutions. Lenders can earn higher returns compared to savings and investment products offered by banks, while borrowers can borrow money at lower interest rates, even after the P2P lending company has taken a fee for providing the match-making platform and credit checking the borrower. The first P2P lending site, Zopa.com, launched in England in Since then many P2P lending sites and platforms have emerged in many countries. In the United States, the two largest P2P lending sites are Prosper and LendingClub. For example, launched in 2006, Proposer has had more than two million members and funded over $2,000 million loans. As the first P2P lending site in China, PPDAI Group Inc.(Paipaidai) has attracted more than five million users, becoming one of the leading P2P platforms (Prosper.com). In this study, we developed online Peer-to-Peer(P2P) lending default prediction models using logistic regression, classification tree algorithms (i.e., CART and C5.0), and multilayer perceptron (i.e., MLP) in order to predict P2P loan default. To verify the feasibility and effectiveness of P2P lending default prediction models, borrower loan data and credit data provided by Lending Club, the biggest United States P2P lending company used in this study. 2. Prior Studies on P2P Lending Since the first P2P platform started in 2005, there has been a growing body of literatures focus on online P2P lending. Many of the studies use tradition data on Prosper, which has made its loan data publicly available. The information of borrowers is divided into hard information and soft information. Hard information includes credit rating, loan amount, ethnicity, gender and so on. Soft information includes social network and social capital of the borrowers. Freedman and Jin(2008) show that the credit rating of the borrowers is positively related to the success rate of loans on Prosper.com. Puro et al.(2010) shows that a lower interest rate decreases the borrower's chances of getting the loan funded, while a smaller loan amount increases the success probability. Herzenstein et al.(2008) show that borrowers financial strength, their listing and publicizing efforts and demographic attributes affect the probability of funding successful. 2

3 Duarte et al.(2012) show that borrowers who have more trustworthy appearance are more likely to have their loans funded. Lin et al.(2013) examined the relationship between online friendship networks and information asymmetry in the largest online P2P lending marketplace, Prosper.com. The results show that online friendships increase the probability of successful P2P funding, lower interest rates on funded loans, and are associated with lower ex post default rates. Larrimore et al.(2011) examined the relationship between language features, trustworthiness, and persuasion success in P2P lending environment. They insist that objective and specific description of loan has positive impact on funding success. Serrano-Cinca et al.(2015) suggested determinant factors of default in P2P lending such as loan purpose, annual income, current housing situation, credit history and indebtedness. Xu et al.(2015) focuses on a specific type of fraud, loan request fraud, which may be unique to lenders on Chinese P2P lending sites due to the lack of nationwide credit rating systems in China. Zhang et al.(2017) conducted by using public dataset from Paipaidai, the largest online P2P lending in China. They suggested that the determinants factors of online P2P lending such as annual interest rate, repayment period, credit grade, successful loan number, failed loan number, gender, and borrowed credit score affect the success rate of P2P loans. Yang et al.(2017) suggested influencing factors of P2P lending success rate based on social capital theory in China. The results show that soft information such as bidding record has a more significant effect on the success rate, and users depend more on the social capital; the bidding records reduce the asymmetry of information, and help increasing the success rate of lending and decreasing the cost of online P2P lending. 3. Experimental Design This study uses loan applications on the Lending Club platform from January 2016 to September 2017 obtained from Lending Club web site ( Lending Club is the biggest US P2P lending company. Among the application set, there are applications obtained the loan in the end (fully paid), and charged off (defaulted). Table 1. Loan Application Status Application status Number of loans Percent Fully Paid % Charged Off(Default) % Total % We divide the influence factors into four aspects, including borrower characteristics, borrower assessment, loan characteristics, and credit information. Borrower characteristics includes annual income, housing situation, and length of employment. Borrower assessment 3

4 includes LC-grade and debt-to-income ratio. Loan characteristics includes loan purpose, loan amount, and interest rate. Credit information includes number of finance trades, delinquency in past 2 years, public records, bank card open to buy, mortgage accounts, and number of bankcard accounts. The value of 1 stands for getting the P2P loan successfully, and 0 means that the loan defaulted. We used the dataset with 14 properties, including annual income, housing situation, length of employment, LC-grade, debt-to-income ratio, loan purpose, loan amount, interest rate, number of finance trades, delinquency in past 2 years, public records, bank card open to buy, mortgage accounts, and number of bankcard accounts. Table 2 shows the variables of the study. Table 2. Variable used in the Study Characteristics Name of variable Description of variable The self-reported annual income provided by the Annual Income Borrower borrower during registration Characteristics Housing Situation Own, rent, mortgage, and other Length of Employment Employment length in years Lending Club categorizes borrowers into seven Borrower LC-Grade different loan grades form A down G Assessment Debt to Income Ratio Borrower s debt to income ratio 14 loan purposes: debt consolidation, credit card, Loan Purpose home improvement, major purchase, medical, car Loan loan, moving, vacation, house, and other Characteristics The listed amount of the loan applied for by the Loan Amount borrower Interest Rate Interest rate on the loan Number of Finance Trades Number of finance trades Delinquency in past 2 Years The number of 30+ days past-due incidences of Credit Information Public Records Bank Card Open to Buy Mortgage Accounts Number of Bankcard Accounts delinquency Number of derogatory public records Bank card open to buy Mortgage accounts The number of bankcard accounts in the borrower s credit file In this study, we used a stepwise logistic regression method, classification tree algorithms (CART and C5.0), and MLP to develop P2P lending default prediction models. Each dataset is split into two subsets: a training set and a validation (holdout) set. The training subset is used to train the prediction models, whereas the validation subset is used to test the model s prediction performance for data that have not been used to develop the classification models. Both the training subset (60% of the larger dataset, with P2P loan data) and the validation subset (40% of the larger dataset, with loan data) are randomly selected. We replicate the entire process of data selection, estimation, and testing five times in order to reduce the impact of random variation in the dataset composition. Cross-validation, a well-known method, is applied to enhance the generalizability of the test results. 4

5 4. Results and Discussion Table 3 compare the prediction performance of logistic regression, CART, C5.0, and MLP using fivefold cross-validation. We can evaluate the prediction performance using the accuracy rate which is calculated by dividing the number of correct predictions by the total number of predictions. Among these models, MLP show the highest average accuracy of 81.78% with the given validation sets, followed by C5.0 with 79.33% and CART with 78.91%. The results from the tests show that the performance of MLP is superior to that of the other classifiers such as logistic regression, CART, and C5.0. MLP always outperform other classifiers in P2P loan default prediction; it can predict borrowers loan default risk more accurately than any other classifier. Table 3. Comparison of Prediction Models Data Set no. Result LR CART C5.0 MLP Data Set 1 Training 61.94% 79.32% 80.34% 82.36% Validation 61.57% 79.21% 79.82% 81.54% Data Set 2 Training 62.08% 79.87% 80.23% 82.22% Validation 61.96% 78.63% 79.40% 82.27% Data Set 3 Training 62.22% 79.70% 80.00% 82.96% Validation 60.79% 78.27% 78.96% 81.34% Data Set 4 Training 61.43% 78.89% 80.13% 81.77% Validation 62.84% 79.14% 79.11% 83.01% Data Set 5 Training 62.12% 79.98% 80.20% 82.99% Validation 60.98% 79.32% 79.37% 80.76% Avg. Training 61.96% 79.55% 80.18% 82.46% Validation 61.63% 78.91% 79.33% 81.78% Note: LR (Logistic Regression), CART (Classification And Regression Trees) C5.0 (C5.0 is significantly faster than C4.5), MLP (Multi-Layer Perceptron) Our study has the following limitations, which require further investigation. First, the results from the study should be generalized. Our study uses only a single selected dataset for system validation. However, only one dataset may not be reliable for making a conclusion. It is necessary to consider a certain number of different datasets for system validation. We believe that other problem domains (bankruptcy prediction, stock market prediction, dividend policy forecasting, and fraud detection) should be investigated in order to generalize the results of this study. Secondly, future research should consider social interaction and herding behavior variables for P2P loan default inputs. References Duarte, J., Siegel, S., Young, L., 2012, Trust and credit: the role of appearance in peer to peer lending. Review of Financial Studies 25(8), Freedman, S. and G.. Z. Jin, 2008, Do social network solve information Problems for Peer-to-Peer lending? Evidence from Prosper.com. Working Paper University of Maryland & NBER, Herzenstein, M., Andrews, R., and Dholakia, U. 2008, The democratization of personal consumer loans? Determinants of success in online peer-to-peer lending communities, Working Paper. Larrimore, L., Jiang, L., Larrimore, J., Markowitz, D., and Gorski, S., 2011, Peer to peer lending: The relationship between language features, trustworthiness, and persuasion success, Journal of Applied Communication Research, 39(1),

6 Lin, M., Prabhala, N. R. and Viswanathan, S., 2013, Judging Borrowers by the Company They Keep: Friendship Networks and Information Asymmetry in Online Peer-to-Peer Lending, Management Science, 59(1), Prosper. Prosper.com: About us (2014), (cited December 30, 2014). Puro, L., Teich, J.E., Wallenius, H., and Wallenius, J., 2010, Borrower decision aid for people-to-people lending, Decision Support Systems, 49, Serrano-Cinca, C., Gutiérrez-Nieto, B., and López-Palacios, L., 2015, Determinants of Default in P2P lending, PloS ONE, 10(10), Xu, Jennifer J., Lu, Y., and Chau, M., 2015, P2P Lending Fraud Detection: A Big Data Approach, Lecture Notes in Computer Science, 2015, 9074, Yang, Z., Zhang Y., and Jia, H., 2017, Influencing Factors of Online P2P Lending Success Rate in China, Annals of Data Science, 4(2), Zhang, Y., Li. H., Hai, M., Li, J. and Li, A., 2017, Determinants of loan funded successful in online P2P Lending, Procedia Computer Science, 122,